Referral Incentives in Crowdfunding

Victor Naroditskiy Sebastian Stein Mirco ToninLong Tran-Thanh Michael Vlassopoulos Nicholas R Jennings

University of Southampton

Abstract

Word-of-mouth, referral, or viral marketing is a highlysought-after way of advertising. In this paper, we in-vestigate whether such marketing can be encouragedthrough incentive mechanisms, thus allowing an or-ganisation to effectively crowdsource their marketing.Specifically, we undertake a field experiment that com-pares several mechanisms for incentivising social me-dia shares in support of a charitable cause. Our exper-iment takes place on a website promoting a fundrais-ing drive by a large cancer research charity. Site visitorswho sign up to support the cause are asked to spread theword about it on Facebook, Twitter or other channels.They are randomly assigned to one of four treatmentsthat differ in the way social sharing activities are incen-tivised. Under the control treatment, no extra incentiveis provided. Under two of the other mechanisms, thesharers are offered a fixed number of points that helptake the campaign further. We compare low and highlevels of such incentives for direct referrals. In the fi-nal treatment, we adopt a multi-level incentive mech-anism that rewards direct as well as indirect referrals(where referred contacts refer others). We find that pro-viding a high level of incentives results in a statisticallysignificant increase in sharing behaviour and resultingsignups. Our data does not indicate a statistically sig-nificant increase for the low and multi-level incentivemechanisms.

Introduction

Within just a few years, we have seen a tremendous riseof online social networks, which have made sharing infor-mation easy and natural. A message that is shared on so-cial networks can reach a large audience without the helpof traditional media outlets. Furthermore, the message maybe targeted towards the right audience, as people sharing itare likely to have friends with similar interests (McPherson,Smith-Lovin, and Cook 2001). A recent social experiment,the DARPA Network Challenge (Defense Advanced Re-search Projects Agency 2010), provided an example of howpowerful this method of sharing information can be: a seem-ingly impossible task of locating 10 red weather balloons

Copyright c© 2014, Association for the Advancement of ArtificialIntelligence (www.aaai.org). All rights reserved.

placed at secret locations throughout the US was solvedwithin 9 hours.

Reaching audiences through social media is particularlyimportant for organisations that do not have a large mar-keting budget. For example, many charitable drives, crowd-funding or crowdsourcing projects (Douceur and Mosci-broda 2007) rely on existing participants to recruit newdonors or members. The Guardian1 reports that 22% of alldonations on a popular charitable giving platform, JustGiv-ing, originate from posts current donors make on Facebook.

In this work, we ask the question of how to effectivelycrowdsource marketing in this way, i.e., how to incentivisethe sharing of a particular message on social media. To thisend, we undertake a natural field experiment where we testthe performance of three incentive mechanisms. The com-pared mechanisms encourage existing users of a website topost a message on Facebook, Twitter or other channels, inorder to attract new users. The mechanisms offer a certainreward for each new signup resulting from a user’s socialmedia post. This reward is expressed in terms of points anddiffers between the various mechanisms as follows:

• 1 extra point for each contact who signs up (low treat-ment),

• 3 extra points for each contact who signs up (high treat-ment),

• 1 extra points for each contact who signs up, 1/2 for eachcontact of a contact, 1/4 for each contact of a contact of acontact and so on (recursive treatment),

• and a base case treatment providing no additional incen-tive (control treatment).

The experiment is run in the domain of charitable crowd-funding, where the goal of sharing a message is to gener-ate support for a charitable cause described on a websitethat we designed. The users that arrive at the website areasked to sign up to “donate a click”. Each click generates 1point that together with the points generated by our referralmechanisms are used to measure the success of the fundrais-ing campaign. Once a threshold of 1,000 points is reached,£500 is donated to support the drive. Once the next target of10,000 points is reached, another £500 is donated.

1http://gu.com/p/38mpx

Proceedings of the Second AAAI Conference on Human Computation and Crowdsourcing (HCOMP 2014)

171

We include both the 1-point and the 3-point mechanismsto test whether the exact amount of points matters for the re-ferral activity. The recursive or multi-level incentive mecha-nism offers stronger incentives than the 1-point mechanism(as it offers the same number of points for direct referrals,but also rewards indirect referrals). However, how it com-pares to the 3-point mechanism depends on a user’s socialnetwork. Indeed, either the recursive or the 3-point mecha-nism may generate more points for the referrer. The recur-sive mechanism is included to see if multi-level incentivesappeal to users. Furthermore, the recursive mechanism re-cently generated a lot of attention as the winning strategy ofthe DARPA Network Challenge (Pickard et al. 2011).

We evaluate mechanisms along three dimensions. First,we look at the effectiveness of each treatment at encourag-ing participants to make referrals, as measured by the num-ber of treatment participants who clicked one of the sharebuttons our website provides. The second dimension is thenumber of referred visitors to the page, and the third one isthe number of those visitors who signed up.

Our main findings are that incentives make a difference inthis context, and the exact amount or magnitude of the incen-tive is important. The 3-point incentive mechanism outper-formed the control treatment on all of the metrics we consid-ered. The recursive incentive mechanism was not any moresuccessful than a simple 1-point mechanism. One might ex-pect that the presence of any kind of incentives to shareshould be enough for most people, while further differentia-tion into 1, 3, or recursive points should have little bearing.However, our results suggest that offering 3 points resultsin more engagement than offering 1 point or compensatingaccording to the recursive mechanism.

The paper is organised as follows. First, we cover relatedwork. Then, we give details of our experimental design andthe execution of our experiment. Finally, we present our re-sults, followed by a discussion of our findings.

Related WorkThere is little existing empirical work on comparing incen-tives for referrals on social media. In a recent study, Castillo,Petrie, and Wardell (2014) also conduct a field experimentin the charitable giving context. The authors describe a fieldexperiment that compared mechanisms offering to donate afixed amount if a user posts about the cause on Facebook.The authors find that offering $1 for posting on Facebookencourages the action. However, a cost-benefit analysis re-veals that under this referral program, each extra dollar do-nated by referred users costs more than $1 in donations bythe company to incentivise referrals. Unlike our study, thereferral compensation money is donated based on the post-ing action, and not based on the impact it generates throughdonations from referred users. Also, our point-based com-pensation does not directly map to money. The points col-lected only result in a donation if a threshold is reached.

The question of referral schemes received a lot of inter-est in the theoretical community. Kleinberg and Raghavan(2005) study a model where incentives must be provided forusers to propagate a question until a node that knows the an-swer is reached, and Cebrian et al. (2012) consider the use

of recursive mechanisms in this context. Naroditskiy et al.(2012) provide a theoretical justification for the recursivemechanism used in the DARPA Network Challenge, anddesirable properties of a referral scheme have been posedin (Douceur and Moscibroda 2007). The question of sybil-proofness of incentive schemes has been considered in a se-ries of papers (Babaioff et al. 2012; Drucker and Fleischer2012; Chen et al. 2013; Lv and Moscibroda 2013). However,these papers concentrate on theoretical issues and do not in-vestigate or compare referral mechanisms empirically withreal users.

Goel, Watts, and Goldstein (2012) analyze diffusion pat-terns in seven online domains for the degree and depthof propagation. They find that viral reach accounts for asmall fraction of users compared to the users reached di-rectly through media or marketing. Furthermore, whateverviral reach occurs is mostly limited to users just one re-ferral away from the users reached directly through media.In other words, referrals through referred audiences are ex-tremely rare. Our findings confirm those results: our referredusers did not generate any successful referrals. While viralreach is dreamed for by many, a more realistic goal is anincreased reach through one-step referrals. Our research ishelping choose the right incentives for this.

This paper contributes to the economics literature on char-itable giving and fundraising, summarized in List (2011).This literature has used field experiments to analyze theeffectiveness of various mechanisms used in fundraising,ranging from incentives schemes like matching grants (Kar-lan and List 2007) to social pressure (DellaVigna, List, andMalmendier 2012) and social information (Shang and Cro-son 2009). In a recent paper, Lacetera, Macis, and Mele(2014) look specifically at online fundraising and show thatthe effect of broadcasting donors’ activities on online net-works is due to homophily rather than social contagion ef-fects. Evidence that personal referrals are important in thefundraising context is presented by Meer (2011). More gen-erally, a number of studies have shown that peer pressurehas an impact on charitable giving (Frey and Meier 2004;Shang and Croson 2005; Smith, Windmeijer, and Wright2014).

There is also a connection to the literature on thresholdpublic goods, i.e., public goods that are provided only ifa stated threshold level of contributions is reached. In oursetting, there is a threshold in terms of points to release acharitable donation and we vary how referrals contribute toreaching the threshold. Related to this point, Cadsby andMaynes (1999) find that a high threshold discourages provi-sion, while high rewards significantly increase contributionsand provision, while Croson and Marks (2000) develop theconcept of step return, defined as the ratio of an individual’svalue of the public good to his share of the cost, and find thata higher step return leads to more contributions.

In a business context, Albuquerque et al. (2012) observethat referral activities of publishers on an online magazineplatform bring significant revenue to the platform. One ofthe recommendations of the paper is to stimulate referral ac-tivity with incentive schemes. Schmitt, Skiera, and Van denBulte (2011) find that referred customers are more valuable

172

to a bank than non-referred customers. A reason for this isthe targeting ability of crowdsourced referrals. Since an ex-isting customer has information about the bank and abouthis friends, he only refers friends whose needs match whatthe bank provides. Domingos and Richardson (2001) pro-vide a framework for estimating the value of each customer,taking into account the social influence that customer has onhis social network.

Our work also complements the growing literatureon crowdfunding, the solicitation of money or resourcesthrough social media to support a particular project (Ger-ber et al. 2014; Belleflamme, Lambert, and Schwienbacher2013). Researchers typically focus on how to efficiently mo-tivate the set of financial supporters (i.e., funders), who arealready in touch with the crowdfunding project, to increasetheir support. In particular, Agrawal, Catalini, and Goldfarb(2011) investigate how one should take into considerationthe geographical distance between the project owners andthe funders. Mitra and Gilbert (2014) study the languageand the phrases used in the description of the project, andtheir impact on the success of the corresponding crowd-funding project. Belleflamme, Lambert, and Schwienbacher(2013) look at the compensations a project owner shouldprovide for funders (e.g., fixed, pre-ordered products, orprofit share), in order to increase the total amount of mone-tary pledges. In a similar way, Ordanini et al. (2014) inves-tigate how project owners should interact with the crowd toimprove the efficiency of the project.

On the other hand, referrals have the potential to signifi-cantly expand not just the set of funders, but also raise thegeneral awareness of a particular project, both of which havebeen identified as key motivators for project owners to par-ticipate in crowdfunding (Gerber and Hui 2013). In relatedwork, the social capital of a project owner, measured bythe number of their Facebook contacts, is shown to be sig-nificantly correlated with the success of a project (Giudici,Guerini, and Rossi Lamastra 2013; Mollick 2014). Throughinterviews, Hui, Gerber, and Gergle (2014) confirm that theefficient exploitation of a project owner’s social network isoften instrumental to the success of a crowdfunding projectand highlight the importance of obtaining endorsement bycontacts that are influential within their own network. In thispaper, we build on these observations and investigate howto effectively encourage such referrals, thereby reaching be-yond the project owner’s direct contacts.

While crowdfunding is a particular application in whichproject owners can benefit from using their social mediacontacts, there have been more general studies on how usersaccess resources within their social networks. In a labora-tory study, Jung et al. (2013) show that the specific type ofsocial capital a user has access to matters — those who re-port that they have some close friends within their networkreceive more responses to a simple favour than those whodo not. Surprisingly, the actual number of friends does notmatter, but those who frequently asked questions in the pastalso receive more responses. The researchers also show thata user’s rhetorical strategy for asking for a favour matters.This study is complementary to our work, as it examines thefactors that contribute to an individual’s success at eliciting

a response from their social network. In contrast, we also ex-amine how to incentivise users to post a message in the firstplace.

In general, posting messages on social networks can incura social cost, as users do not want to waste the time or effortof others and because they often care about how they are per-ceived (Sleeper et al. 2013). Rzeszotarski and Morris (2014)measure this social cost explicitly by offering users a mon-etary reward for asking their followers on Twitter to answerquestions. They show that a higher monetary reward leadsto a higher probability of users asking these questions. How-ever, in this work, we consider referrals rather than genericquestions, and we examine non-monetary incentives.

Field ExperimentIn order to compare a range of mechanisms for incentivisingreferrals, we partnered with Cancer Research UK (CRUK),one of the UK’s largest charitable organisations, to carry outa natural field experiment. To this end, we set up a websiteto raise awareness of one of CRUK’s existing fundraisingcampaigns. The website provided an indirect incentive forvisitors to engage with it by adding a point to an overalltotal displayed on the website whenever a new user signedup. This total was then used to release donations from anoutside donor to the CRUK campaign — whenever certainpoint thresholds were reached, a new donation was made.

After signing up, users were randomly assigned to one ofa set of referral mechanisms. Here, some of the users wereable to add further points to the common total by success-fully referring their social contacts to the website. This al-lowed us to test whether such additional incentives can in-crease the probability of a successful referrals, and, by test-ing various mechanisms, whether the magnitude of the re-ward is significant and whether recursive schemes offer abenefit in this setting.

In the following, we describe the setup of our field exper-iment and the tested referral mechanisms in more detail.

Website

Our website, http://outruncancer.co.uk/, was set up inSeptember 2013 in consultation with CRUK and providedvisitors with information about oesophageal cancer anda particular CRUK fundraising campaign, “The CancerMarathon”. This campaign was initiated by six medics atthe Southampton General Hospital, who were participatingin the New York Marathon in November 2013 to raise moneyfor research on oesophageal cancer.

The stated purpose of our website was to support the cam-paign by spreading awareness about it and thereby attractfurther donations from the public. To achieve this and en-courage engagement with our site, visitors could contributeone point to a total amount prominently displayed on thewebsite simply by signing up. This did not incur a financialcost for them and involved either entering an email addressand password, or signing in through existing Facebook orTwitter accounts. Importantly, we pledged to make severaldonations of £500 each to CRUK’s fundraising campaign,as soon as certain thresholds in the total number of pointsgenerated by all users were reached:

173

Figure 1: Main panel of the landing page.

1. £500 when 1,000 points were reached in total

2. £500 when 10,000 points were reached in total

3. £500 when 50,000 points were reached in total

4. £500 when 100,000 points were reached in total

We chose this scheme, in order to cope with varying to-tal amounts of visitors to our site. The first threshold wasrelatively low, in order to be able to generate at least one do-nation, while the next thresholds were spaced increasinglyfurther apart, to ensure that the experiment did not have toterminate due to running out of money for donations.

Figure 1 shows the main part of the initial landing page(the full page is shown in Figure 9 in the appendix). Here,a prominent box displayed the current progress to the nextdonation, a summary of how the website works and a regis-tration panel. Below this (see Figure 9), there were expand-able sections offering more details about the campaign andthe website. Importantly, as the experiment was designed asa natural field experiment, the website did not disclose that(anonymised) data from users would support research on re-ferral incentives. This was crucial, because knowledge of theresearch could have altered the users’ referral behaviour.

Apart from the dynamically updated total number ofpoints, all visitors to the website received exactly the sameinformation, and there was no explicit mention of referrals.This was done to avoid any priming of the visitors, and tomitigate self-selection bias regarding referral behaviour. Theonly varying parameter – the total number of points – did notaffect our results as we only used observations from the peo-ple who participated after the first target of 1,000 points hadbeen reached and when the next target of 10,000 was very far

away (the experiment ended with the total number of pointsjust over 2,000).

Referral Mechanisms

Once a visitor signed up to the website, they were presentedwith a page thanking them for their contribution and sug-gesting a number of ways they can help the fundraising cam-paign (Figure 2 shows the main part of this page, while thefull page is shown in Figure 10 in the appendix). In par-ticular, they were asked to invite their friends via a rangeof channels: by copying and sending a personalised link, bysharing a link on Facebook or Twitter, or by sending an emailwith the link.

Here, we provided some users with additional incentives,depending on the treatment they were randomly allocated to.This random allocation took place on signup and used a sim-ple blocked randomisation scheme, whereby the allocationmechanism repeatedly cycled through random permutationsof the available treatments (thus ensuring a near-balancedallocation to treatments).

The treatments we considered were as follows:• Control: Users on this treatment were offered no addi-

tional referral incentives.• Low: Users on this treatment were told that 1 additional

point would be added to the total when they successfullyreferred someone else to the website (in addition to thesignup point generated by that contact).

• High: Similar to Low, but users were offered 3 additionalpoints.

• Recursive: Users were told that 0.5n−1 additional pointswould be added to the total when they successfully re-

174

Figure 2: Referral options after signing up (control treatment shown).

ferred someone else to the website, including through in-direct referrals, where n is the distance to the user.2 Thus,if the user successfully referred another contact, 1 addi-tional point would be added to the total. If that referredcontact successfully referred someone else, 0.5 additionalpoints would be added, and so on.The wordings of these treatments, as shown to the users,

are summarised in Figure 3. Referrals were tracked througha unique ID that was allocated to each user and then includedin the referral links generated through the four referral meth-ods on the website (direct link, Facebook, Twitter or email).These links also encoded which of the four methods wasused, allowing us to track, for each referred user, who theywere referred by and through which referral method. Over-all, we recorded the following three key metrics for eachuser:• Number of channels used: How many of the four referral

methods were initiated by the user on our website. Werecord this when the user clicks on one of the buttons (orthe referral link) shown after signing up.3

• Number of resulting visits: How many visits were sub-sequently generated by people following the user’s uniquereferral link (but did not necessarily sign up).

• Number of resulting signups: How many new users sub-sequently signed up through the user’s referral link (this2We say user A indirectly referred B, if B was not (directly)

referred by A, but by another user C, which was either referred byA or is also an indirect referral of A. The distance between A andB is the length of the resulting referral chain (a direct referral hasdistance 1).

3Note that this is not an exact measure of referral messagesposted, as users may still cancel the message on the social mediawebsite before posting or may never share the link.

is number of the user’s direct referrals).Note that for all four treatments, there is an implicit incen-

tive for users to refer contacts to the website, as all signupsgenerate 1 point towards the total. The low, high and recur-sive treatments add an additional incentive on top of this.The first two of these, low and high, are simple schemes thatwe chose to investigate whether the magnitude of this in-centive is relevant — in the first case, it is equivalent to theoriginal incentive offered to the user for signing up; in thesecond case, it is significantly higher than the original in-centive. We included the recursive scheme, because it hasbeen used successfully in previous crowdsourcing applica-tions (as described in Related Work) and because it mayencourage users to specifically target contacts with a widesocial reach. From a user’s perspective, it dominates the lowscheme, because it also awards 1 additional point for eachdirect referral, but may generate further points through sub-sequent indirect referrals. However, the relative attractive-ness of recursive and high depends on the user’s social net-work. For example, a user who has only one friend whois very well-connected, prefers the recursive scheme as hewill be rewarded half a point on each signup of the well-connected friend’s friends.

Users were able to log back into the website at any time toview how many of their contacts had signed up and (exceptfor the control treatment) how many additional points hadbeen generated through this. We also gave users the optionto donate directly to the Cancer Marathon campaign, whichwe matched one-to-one (up to £10 per user).

Subject Recruitment

The website went live in September 2013 and over the nexttwo months, 1,577 unique users signed up, generating a to-tal of 2,051 points (one point was generated on signup; the

175

(a) low treatment (b) recursive treatment

Figure 3: Explanatory texts for low and recursive treatments. Text for high treatment is identical to low except for the numberof additional points (3).

Figure 4: Email to invite participants.

remaining 474 points were generated for successful refer-rals) and an additional £418 in user donations. As a resultof meeting the first threshold, we donated £500, as well as£408 in matching donations (this is slightly lower than thetotal amount donated by the users due to the £10 limit peruser).

Initial users were recruited from a range of sources. Somepublicity was generated by CRUK and the medics run-ning the Cancer Marathon, including through Twitter mes-sages and a local radio interview. However, the bulk of theusers were attracted through emails within the University ofSouthampton, sent to staff and student mailing lists. Thisemail did not mention that the website was part of a researchproject and did not identify the academics involved in it (seeFigure 4).

During the course of the experiment, we received feed-back from users, indicating that the points mechanism wasnot explained clearly and succinctly enough. For this reason,we made a number of changes to the website in mid-October2013 — primarily to highlight the current points collected todate at the top of the page and to emphasise their purpose. As

the experimental conditions changed at that time, we focusonly on the 412 participants recruited after the change in theremainder of this paper. These are mainly students from theFaculties of Humanities, Social Sciences, and Health Sci-ences and their referred contacts, as we emailed their inter-nal mailing lists after the change (and had not previouslycontacted them). Changing the experiment half-way did notpresent any complications as far as the analysis of the re-sults goes. Restarting midway can be viewed as starting theexperiment at the midpoint as the first part was simply ig-nored and there was no contamination from it.

After the redesign, we contacted 17,686 University ofSouthampton students via email. There were 1,042 uniquevisits that resulted in 412 signups. In terms of the effective-ness of various social media channels, Twitter was respon-sible for generating 55% the referred traffic while Facebookbrought 36% of referred page visits. However, the number ofpage visits that turned into signups is much higher for Face-book: 62% compared to Twitter’s 8%. The remaining visitsand signups were generated by sharing a link without the useof the social media buttons.

There were too few donations to perform meaningful sta-tistical analysis of differences across treatments, but we re-port them here for completeness. After the redesign, the to-tal of £40 was donated by 6 users. Five donors were in theHigh treatment donating the total of £30 and one user whodonated £10 was in the Low treatment.

Results

Table 1 summarises the total number of times all channelswere used under each treatment, the number of resulting vis-its, and the number of resulting signups. We observe that the3-point treatment resulted in higher totals across the board.There is no other clear dominance across treatments. Some-what of an outlier is a low ratio of recursive treatment’s vis-its and signups. There were 103 visits but only 2 signupsresulting in a 2% signup ratio. The corresponding ratios forthe control, low, and high treatment are 14%, 6%, and 11%respectively. Given the limited number of observations, wecould not test for whether there is an association betweenratio and treatment.

We present details of visits and signups in Figures 5 and 6.The histogram in Figure 6 shows how many referrers signedup exactly 1 friend, 2 friends, . . . , 8 friends. For the high

176

Control Low High Recursive TotalChannels used 13 15 32 28 88Resulting visits 37 98 218 103 456Resulting signups 5 6 24 2 37

Table 1: Total outcomes per treatment.

Figure 5: Histogram of visits generated per user.

treatment, there were 6 referrers who signed up 1 friend, 3referrers who signed up 2 friends, 1 referrer who signed up4, and 1 referrer who signed up 8 friends. The histogramof visits in Figure 5 should be read in a similar manner: thefirst set of bars shows how many referrers generated between1 and 7 referred visits, the second set of bars shows howmany referrers generated between 8 and 14 referred visits,etc. For example, between 8 and 14 visits were generated byzero referrers in the control treatment, by 1 referrer in thelow treatment, by 3 referrers in the high treatment, and by 5referrers in the recursive treatment.

Next, we evaluate the results we obtained from our ex-periment. In particular, we focus on both the extensive andintensive margins of the behaviour under investigation. Theextensive margin uses binary (0/1) indicators that reflectwhether a certain success measure is achieved. In our case,relevant indicators are whether or not a referrer generated avisit and whether or not she generated a signup. The inten-sive margin quantifies the degree of success by looking atthe number of visits and signups generated by each referrer.

Extensive Margin

In this section we examine whether there are differences inthe outcomes of interest in the extensive margin. We fo-cus on three binary outcomes: (i) whether a referral chan-nel was taken; (ii) whether at least one visit was generated;(iii) whether at least one signup was generated. All resultsregarding these three outcomes are summarised in Table 2.

With regards to referral channels, we see in the first sec-

Figure 6: Histogram of signups per user.

tion of Table 2 that people in our sample used some of thetools we provided to refer other potential participants in 16%of the cases (65 of the total 412). With no incentives, thishappened in only 10% of the cases, while the percentage isalways higher when referral incentives are present. In par-ticular, the high incentive treatment is successful in elicitinga referral through some channel among 21% of the people,while for the low incentive treatment the corresponding fig-ure is 14%. The recursive treatment is in between, with afigure of 18%. The p-value of a chi-squared test is, at 0.106,which does not allow us to reject the null hypothesis of in-dependence between the likelihood of undertaking some re-ferral channel action and the various treatments.

Out of the 16% of people who used some of the refer-ral tools, 75% were successful in generating some visits tothe website. That is, out of 65 people who used a referralchannel, 46 resulted in at least one new user visiting thewebsite (see the second section of Table 2). The effect ofincentives in generating additional traffic is even more evi-dent, with only 6% of the people in the control group gen-erating some visits, while the figure for the high incentivegroup, at 18%, is three times higher. Again, the low incen-tives are less successful (9%) and recursive incentives arepositioned in between (12%). Figure 7 displays the percent-ages along with 95% confidence intervals. A chi-squared testrejects the null hypothesis of independence between the like-lihood of generating some visits and the various treatments(p-value=0.028).

The percentage of people who generated additional

177

Control Low High Recursive TotalAt Least One Channel Used?

Yes 10 (9.8%) 14 (13.5%) 22 (21.4%) 19 (18.4%) 65 (15.8%)No 92 (90.2%) 90 (86.5%) 81 (78.6%) 84 (81.6%) 347 (84.2%)

Resulting Visits?

Yes 6 (5.9%) 9 (8.7%) 19 (18.4%) 12 (11.7%) 46 (11.2%)No 96 (94.1%) 95 (91.3%) 84 (81.6%) 91 (88.3%) 366 (88.8%)

Resulting Signups?

Yes 2 (2%) 5 (4.8%) 11 (10.7%) 1 (1%) 19 (4.6%)No 100 (98%) 99 (95.2%) 92 (90.3%) 102 (99%) 393 (95.4%)

Total Participants 102 104 103 103 412

Table 2: Referral channels, visits and signups per treatment (extensive margin).The p-value of a chi-squared test for independence is 0.106, 0.028 and 0.004 for channels used, visits and signups, respectively.

Figure 7: Percentage of participants whose referrals broughtat least one website visitor. 95% confidence intervals areshown.

signups is 4.6% (see the third section of Table 2). Again,our referral treatments seem to be effective in boosting thisnumber. Only 2% of people assigned to the control groupgenerated additional signups, while in the high incentivescase, the figure is much higher, at 11%. Low incentives areless effective (5%), while the recursive treatment is the leasteffective (1%). Figure 8 displays the percentages along with95% confidence intervals. Again, a chi-squared test rejectsthe null hypothesis of independence between the likelihoodof generating additional signups and the various treatments(p-value=0.004).

These results are confirmed by a regression analysis (Ta-ble 3), where it is evident that the high incentives treatmentis effective in generating significantly more referral actions,more visits, and more signups than the control group.4 For

4We present regression analysis here and not ANOVA becauseit allows for pairwise comparisons between each of the treatmentsand the control, which is our primary interest.

Figure 8: Percentage of participants whose referrals resultedin at least one signup. 95% confidence intervals are shown.

the other treatments, the coefficients are generally positivebut not statistically significant. Note also that with regardsto signups, the difference between the high incentives andthe recursive treatment is also statistically significant (p-value=0.007), while the difference between high and low ismarginally not (p-value=0.116). Finally, for visits, we finda significant difference between the coefficients of the highversus the low treatment (p-value=0.04).

The model we use for the analysis is a standard probit,where we fit the following equation: Pr(signups = 1) =Φ(b0 + b1 ∗ High + b2 ∗ Low + b3 ∗ Recursive). Φ is thecumulative normal distribution. High is a variable taking thevalue of 1 if observation is in High treatment, 0 otherwise.Low is a variable taking the value of 1 if observation is inLow treatment, 0 otherwise. Recursive is a variable takingthe value of 1 if observation is in Recursive treatment, 0 oth-erwise. The omitted treatment is Control and, therefore, forthat treatment we report the coefficient on the constant, b0.The asterisks indicate the significance of tests on whether

178

Control Low High RecursiveReferral channels -1.293*** (0.17) 0.188 (0.23) 0.499** (0.22) 0.394* (0.22)Visits -1.565*** (0.20) 0.202 (0.27) 0.666*** (0.25) 0.372 (0.26)Signups -2.062*** (0.29) 0.398 (0.36) 0.818** (0.33) -0.276 (0.47)

Table 3: Regression analysis of the treatments.Probit regression — robust standard errors are given in parentheses. *** [**] (*) denote significance at 1, [5], (10) % level.

the coefficients are different from 0.

Intensive Margin

In this section we examine whether there are differences inthe outcomes of interest in the intensive margin. We focusagain on three outcomes: (i) number of channels used; (ii)total number of visits generated; (iii) total number of signupsgenerated (as described in the Referral Mechanisms Sec-tion).

Table 4 presents averages for the various outcomes acrossthe four treatment groups. It is evident that the high incen-tives treatment achieves the highest average in all measures,and in particular, with regards to the number of signups.In fact, t-tests of differences in means of the various out-comes between participants in the high and the control groupindicate statistically significant differences in all cases (p-values are 0.008 for channels, 0.018 for visits and 0.069 forsignups).

Finally, when we perform nonparametric tests (Mann-Whitney test) of differences in the distribution of the vari-ous outcomes between participants in the high and the con-trol group we find statistically significant differences in allcases (p-values are 0.018 for channels, 0.005 for visits and0.012 for signups). When comparing high to low treatment,we find significance differences for visits (p=0.04), whilefor sign-ups the p-value is 0.11. Also, the difference in thedistribution of sign-ups of high and recursive treatments issignificant (p-value=0.003).

Discussion

Word-of-mouth, referral, or viral marketing is a highlysought-after way of advertising. Ideally, it would occur or-ganically because people find the promoted message of highvalue. It is however notoriously difficult to design viral mes-sages. In this paper we attempt not to optimise the message,but to find out how sharing of a given message can be stim-ulated.

Our finding that the 3-point mechanism is more effectivethan the 1-point mechanism points to the need to choose thelevel of compensation carefully. This is contrary to the intu-ition that it is not the exact level, but rather the presence ofsome form of incentives, which has the most effect. Consis-tent with this intuition, Castillo, Petrie, and Wardell (2014)find little difference in referral activity between a mecha-nism that pays $1 and $5 for a sharing action. The corre-sponding sharing rates are 37% and 39%.

One might expect incentives to not matter at all in the con-text where points do not directly correspond to donations,but only matter if a sufficient number of points is reached.This is even more plausible if the threshold number of points

does not appear to be within easy reach. This was the casein our experiment. All of the observations used in the dataanalysis came from the users who signed up after the firsttarget of 1,000 points was reached, and when the second tar-get of 10,000 points was far away. Indeed, the total numberof points at the end of the experiment was just above 2,000.No user could reasonably expect that their friends wouldbring 8,000 points no matter which treatment they were partof. The result that users still reacted to the 3-point treatmentmore than to the 1-point or recursive treatments is even moreinteresting given this context.

Our mechanisms reward successful referrals and they arecomplementary to the mechanisms that reward referral ac-tions (i.e., simply sharing a message about the cause, re-gardless of whether this attracts new signups). Cost-benefitanalysis of paying for referral actions results in cautionarylessons (Castillo, Petrie, and Wardell 2014). We conjecturethat the mechanisms considered in our work can be morecost-effective. In particular, one could guarantee an arbitrarycost-benefit ratio by offering the reward as a fraction of theactual donation made by the referred donors.

From an organisation’s point of view, encouraging socialmedia shares is crowdsourcing marketing activities. Takinga broad view on this line of research, we are after learn-ing how to encourage existing participants or customers todo marketing for the project or business. In future work,we plan to investigate this question of how to balance thebenefits of an incentive scheme with its related costs. Wewill also explore whether gamification mechanisms, includ-ing leaderboards or badges, are effective in this context, andwe will carry out further experiments in other domains, suchas crowdsourcing projects and commercial websites.

As in any empirical research, and in particular in an ex-perimental setting, there is a concern about generalizability.Do our findings apply beyond the charitable context? Wouldwe find similar results with a different population or are uni-versity students somehow special, for instance because theirsocial network has specific characteristics? The fact that weuse a control group implies that differences between studentsand non-students related, for instance, to the average num-ber of friends or the likelihood to use social networks do notmatter. Nor do differences, to make another example, in thegeneral propensity to respond to a charitable versus a noncharitable cause. What our results show is a differential re-sponse to the various incentive mechanisms we test and toassess the robustness of this finding in different contexts andwith different populations we are currently conducting fur-ther studies.

179

Control Low High Recursive TotalAction channels used 0.1 0.14 0.26 0.23 0.18Resulting visits 0.36 0.94 2.12 1 1.11Resulting signups 0.05 0.06 0.23 0.05 0.09

Table 4: Average outcomes per treatment.

Acknowledgements

We would like to thank the team at CRUK and, in particular,Hannah Fox and Tim Underwood. This work was funded bythe Strategic Interdisciplinary Research Development Fundof the Faculty of Social and Human Sciences at the Uni-versity of Southampton, and by the EPSRC ORCHID pro-gramme grant. This work is also supported by the BritishAcademy Small Research Grant SG131130.

References

Agrawal, A. K.; Catalini, C.; and Goldfarb, A. 2011. Thegeography of crowdfunding. National Bureau of EconomicResearch.Albuquerque, P.; Pavlidis, P.; Chatow, U.; Chen, K.-Y.; andJamal, Z. 2012. Evaluating promotional activities in an on-line two-sided market of user-generated content. MarketingScience 31(3):406–432.Babaioff, M.; Dobzinski, S.; Oren, S.; and Zohar, A. 2012.On bitcoin and red balloons. In Proceedings of the 13thACM Conference on Electronic Commerce, 56–73. ACM.Belleflamme, P.; Lambert, T.; and Schwienbacher, A. 2013.Crowdfunding: Tapping the right crowd. Journal of BusinessVenturing.Cadsby, C. B., and Maynes, E. 1999. Voluntary provi-sion of threshold public goods with continuous contribu-tions: experimental evidence. Journal of Public Economics71(1):53–73.Castillo, M.; Petrie, R.; and Wardell, C. 2014. Fundraisingthrough online social networks: A field experiment on peer-to-peer solicitation. Journal of Public Economics 114(0):29– 35.Cebrian, M.; Coviello, L.; Vattani, A.; and Voulgaris, P.2012. Finding red balloons with “split” contracts: robustnessto individuals’ selfishness. In ACM Symposium on Theory ofComputing (STOC).Chen, W.; Wang, Y.; Yu, D.; and Zhang, L. 2013. Sybil-proof mechanisms in query incentive networks. In Proceed-ings of the fourteenth ACM conference on Electronic com-merce, 197–214. ACM.Croson, R. T., and Marks, M. B. 2000. Step returns in thresh-old public goods: A meta-and experimental analysis. Exper-imental Economics 2(3):239–259.Defense Advanced Research Projects Agency. 2010.DARPA Network Challenge Project Report.DellaVigna, S.; List, J. A.; and Malmendier, U. 2012. Test-ing for altruism and social pressure in charitable giving*.The Quarterly Journal of Economics.

Domingos, P., and Richardson, M. 2001. Mining thenetwork value of customers. In Proceedings of the Sev-enth ACM SIGKDD International Conference on Knowl-edge Discovery and Data Mining, KDD ’01, 57–66. NewYork, NY, USA: ACM.Douceur, J. R., and Moscibroda, T. 2007. Lottery trees:motivational deployment of networked systems. In Murai,J., and Cho, K., eds., SIGCOMM, 121–132. ACM.Drucker, F. A., and Fleischer, L. K. 2012. Simpler sybil-proof mechanisms for multi-level marketing. In Proceedingsof the 13th ACM conference on Electronic commerce, 441–458. ACM.Frey, B. S., and Meier, S. 2004. Social comparisons and pro-social behavior: Testing” conditional cooperation” in a fieldexperiment. The American Economic Review 94(5):1717–1722.Gerber, E. M., and Hui, J. 2013. Crowdfunding: Motiva-tions and deterrents for participation. ACM Transactions onComputer-Human Interaction (TOCHI) 20(6):34.Gerber, E. M.; Muller, M.; Wash, R.; Irani, L. C.; Williams,A.; and Churchill, E. F. 2014. Crowdfunding: an emergingfield of research. In 32nd Annual ACM Conference on Hu-man Factors in Computing Systems (CHI 2014), 1093–1098.ACM.Giudici, G.; Guerini, M.; and Rossi Lamastra, C. 2013. Whycrowdfunding projects can succeed: The role of proponentsindividual and territorial social capital. Available at SSRN2255944.Goel, S.; Watts, D. J.; and Goldstein, D. G. 2012. The struc-ture of online diffusion networks. In Proceedings of the 13thACM Conference on Electronic Commerce, EC ’12, 623–638. New York, NY, USA: ACM.Hui, J.; Gerber, E.; and Gergle, D. 2014. Understanding andleveraging social networks for crowdfunding: implicationsfor support tools. In CHI’14 Extended Abstracts on HumanFactors in Computing Systems, 2083–2088. ACM.Jung, Y.; Gray, R.; Lampe, C.; and Ellison, N. 2013. Favorsfrom facebook friends: unpacking dimensions of social cap-ital. In Proceedings of the SIGCHI Conference on HumanFactors in Computing Systems, 11–20. ACM.Karlan, D., and List, J. A. 2007. Does price matter in chari-table giving? evidence from a large-scale natural field exper-iment. The American economic review 97(5):1774–1793.Kleinberg, J., and Raghavan, P. 2005. Query incentive net-works. In Foundations of Computer Science, 2005. FOCS2005. 46th Annual IEEE Symposium on, 132–141. IEEE.Lacetera, N.; Macis, M.; and Mele, A. 2014. Viral altruism?generosity and social contagion in online networks.

180

List, J. A. 2011. The market for charitable giving. TheJournal of Economic Perspectives 157–180.Lv, Y., and Moscibroda, T. 2013. Fair and resilient incentivetree mechanisms. In Proceedings of the 2013 ACM sym-posium on Principles of distributed computing, 230–239.ACM.McPherson, M.; Smith-Lovin, L.; and Cook, J. M. 2001.Birds of a feather: Homophily in social networks. AnnualReview of Sociology 27:415–444.Meer, J. 2011. Brother, can you spare a dime? peer pressurein charitable solicitation. Journal of Public Economics 95(7-8):926 – 941.Mitra, T., and Gilbert, E. 2014. The language that getspeople to give: Phrases that predict success on kickstarter.CSCW.Mollick, E. 2014. The dynamics of crowdfunding: An ex-ploratory study. Journal of Business Venturing 29(1):1–16.Naroditskiy, V.; Rahwan, I.; Cebrian, M.; and Jennings,N. R. 2012. Verification in referral-based crowdsourcing.PLoS ONE 7(10).Ordanini, A.; Miceli, L.; Pizzetti, M.; and Parasuraman,A. 2014. Crowd-funding: transforming customers into in-vestors through innovative service platforms. Journal of Ser-vice Management 22(4):443–470.Pickard, G.; Pan, W.; Rahwan, I.; Cebrian, M.; Crane, R.;Madan, A.; and Pentland, A. 2011. Time-critical social mo-bilization. Science 334(6055):509–512.Rzeszotarski, J. M., and Morris, M. R. 2014. Estimatingthe social costs of friendsourcing. In Proceedings of the32nd annual ACM conference on Human factors in comput-ing systems, 2735–2744. ACM.Schmitt, P.; Skiera, B.; and Van den Bulte, C. 2011. Re-ferral programs and customer value. Journal of Marketing75(1):46–59.Shang, J., and Croson, R. 2005. Field experiments in chari-table contribution: The impact of social influence on the vol-untary provision of public goods. Unpublished manuscript.Shang, J., and Croson, R. 2009. A field experiment in char-itable contribution: The impact of social information on thevoluntary provision of public goods. The Economic Journal119(540):1422–1439.Sleeper, M.; Balebako, R.; Das, S.; McConahy, A. L.; Wiese,J.; and Cranor, L. F. 2013. The post that wasn’t: exploringself-censorship on facebook. In Proceedings of the 2013conference on Computer supported cooperative work, 793–802. ACM.Smith, S.; Windmeijer, F.; and Wright, E. 2014. Peer effectsin charitable giving: Evidence from the (running) field. TheEconomic Journal n/a–n/a.

181

Figure 9: The landing page shown to new visitors.

182

Figure 10: Call to action after signing up (control treatment shown).

183